GFS

Goals (shared storage)

• capacity
  • e.g., 1000 servers, 300TB
• performance
• fault tolerance
  • map-reduce

Approach

• filesystem-like API
  • proprietary library (write/read/append)
  • not POSIX
• single master (metadata)
  • (filename, offset) -> chunk
• chunk size: 64MB. why not smaller, such as 4KB(hdd), 4MB (ssd)?
• 3-way replication

How GFS works

• client communicates with master to retrieve data locations
• client buffers information locally
• client sends to data servers

Performance

• why single master and 64 MB sufficient?
  • workload: large files; sequential read/write
• not a good design if
  • small files (aggregate)
  • random accesses (buffer)

Consistency

• correctness: outcome = expectation
  • concurrency
  • failures
• tradeoff
  • weak consistency: easier to implement, hard to use
  • strong consistency: hard to implement, easy to use

Case 1 (strawman, inconsistent with concurrency)

S1: C1  C2
S2: C2  C1

Case 2 (consistent with concurrency)

S1(P): C1 C2 C1-id C2-id      
S2   : C2 C1             S1-id-C1 S1-id-C2

Case 3 (consistent but undefined)

• a write operation breaks into many smaller write operations

Case 4 (inconsistent with failures)

• follower fails
• primary fails
• two primary? (what about serial numbers?)
  • leases to avoid two masters

Case 5 (more consistency anomaly)

S1(P): C1 C2 C2-id C1-id C3-read
S2   : C2 C1                     S1-id-C2 C3-read S1-id-C1 C3-read